Re: Graphics Card Failure Rates

* First of One:
> Interesting failure statistics from a "large French e-retailer" covering
> parts sold between September 2008 and March 2009.

[snip]

> PNY seems to be doing quite well.

I don't think this statistics is really related to hardware reliability.
While PNY indeed does produce good cards (they are also the only
boardpartner for Quadro professional gfx cards), it is probably more
related to what the buyers do. For example, PNY hardly is the brand
choosen by gamers that want to overclock. Those that show the high
return rates on the other sides are brands that aim on the "enthusiast"
market (gamers) with factory-overclocked products. Such products are
probably overclocked even more, and often also modified (i.e. by
replacing the cooler).

The chipset statistics is equally useless. There very likely is no
relaibility difference between a GTX 260 or GTX 280 GPU itself, however
differences in reliability often are due to different operating
conditions (i.e. clock rates, temperaures etc. Therefore it is silly to
compare chipsets without telling under what conditions (i.e. what gfx
card model, what clock rate, what temperatures etc) they ran.

At the end of the day it is just what it is, a return statistics of a
certain french retailer. It however says nothing about the reliability
of a specific gfx card brand or chipset.

Benjamin

* First of One:

>> For example, PNY hardly is the brand choosen by gamers that want to
>> overclock. Those that show the high return rates on the other sides are
>> brands that aim on the "enthusiast" market (gamers) with
>> factory-overclocked products. Such products are probably overclocked even
>> more, and often also modified (i.e. by replacing the cooler).
>
> That is one school of thought. The other school of thought says the Golden
> Samples are purchased by "armchair" enthusiasts who like the idea of
> overclocking, but prefer to let the factory do it for them under warranty.

This is probably very rare. Either someone is interested in overclocking
(in which case he wants to push the card to the limit) or he isn't (in
which case he wouldn't pay the premium for a bit of overclocking).

> The return rates more likely indicate that despite Gainward's best efforts
> with stress screening, the factory overclocking did result in slightly more
> failures... which may be okay for Gainward, because its profit margins on
> the overclocked cards are higher.

Maybe, but it is even more likely that the cards have been returned
because the buyer realized that the high premium for factory
overclocking was a waste of money.

However, without more details (why the cards have been returned) we can
only guess. And the lack of details makes these statistics even more
worthless.

>> The chipset statistics is equally useless. There very likely is no
>> relaibility difference between a GTX 260 or GTX 280 GPU itself,
>
> The GTX 260 and 280 share the same GT200 core, same cooler, but the GTX 280
> generates more heat even at reference clocks. On what basis do you think
> there is no reliability difference between the two GPU tiers?

First, simply because as you say the GPU itself is identical, and the
differences in heat dissipation and power consumption are not
groundbreaking. And second, my last employer was Nvidia customer with
access to many details that only board partners do have, and no-one is
aware of any general reliability differences between GTX 260 and GTX 280.

>> At the end of the day it is just what it is, a return statistics of a
>> certain french retailer. It however says nothing about the reliability
>> of a specific gfx card brand or chipset.
>
> On the contrary, the return statistics take a real-world cross section of
> current products. If you are shopping for a card, you'll trip over all kinds
> of overclocked cards. Hell, AMD claims to have lowered the reference clocks
> on some cards (e.g. HD4890), just so board vendors can artificially create
> "overclocked" SKUs. With a sufficiently large sample size (as it was in this
> case), the average return rate per brand or chipset is representative of
> what a buyer may see.

Maybe. Maybe not. Who knows. Fact is, with a survey created by a single
retailer which doesn't tell you side parameters you don't know ****. For
example, you don't know why cards have been returned. In the EU, online
retailers are required to take back an item within a certain period of
time after purchase if the customer doesn't like it anymore, so it isn't
known how much of the cards actually had defects and haven't been given
back because the customer was disappointed by the performance, regreted
the buy because the next day the cheaper/faster/cooler successor came
out, or because he found a better deal elsewhere. You also don't know if
the statistics only cover retail products or also bulk/OEM versions, or
if it includes grey market versions that were not made for the European
market. You don't know if the same return rates are seen by other "large
French e-retailers", or retailers in other countries, so you don't know
if this statistic from a single retailer really reflects average return
rates, or for some reason shows quite unusual return rates, i.e. maybe
because the shop tries to sell products that have been returned on to
other customers to avoid having to return them to his distributor
(something which isn't rare in the retail market), which means that a
single defective card would be counted multiple times as "return".

This survey contains so little information that no sane and logic
conclusions can be drawn from it. Of course, you could tell yourself
that this shows the defect rates of various products, but that is
exactly as worthless as this survey.

Benjamin